PaperlayerView exact revision on arXiv

Source-linked AI summary

Rethinking Modern Communication from Semantic Coding to Semantic Communication

Kun Lu, Qingyang Zhou, Rongpeng Li, Zhifeng Zhao, Xianfu Chen, Jianjun Wu, Honggang Zhang

arXiv:2110.08496v3eess.SP

TL;DR

Bit-focused communication can waste bits and fail to convey intended meaning, while existing semantics-aware approaches leave channel variation and semantic objectives insufficiently addressed. The paper proposes confidence-guided JSNC and reinforcement-learning semantic communication; evaluations report reduced bit requirements and meaning-focused outcomes. It also identifies implementation cost as a concern for repeated distillation.

Abstract

from arXiv · show

Modern communications are usually designed to pursue a higher bit-level precision and fewer bits while transmitting a message. This article rethinks these two major features and introduces the concept and advantage of semantics that characterizes a new kind of semantics-aware communication framework, incorporating both the semantic encoding and the semantic communication problem. After analyzing the underlying defects of existing semantics-aware techniques, we establish a confidence-based distillation mechanism for the joint semantics-noise coding (JSNC) problem and a reinforcement learning (RL)-powered semantic communication paradigm that endows a system the ability to convey the semantics instead of pursuing the bit level accuracy. On top of these technical contributions, this work provides a new insight to understand how the semantics are processed and represented in a semantics-aware coding and communication system, and verifies the significant benefits of doing so. Targeted on the next generation's semantics-aware communication, some critical concerns and open challenges such as the information overhead, semantic security and implementation cost are also discussed and envisioned.

I. INTRODUCTION

Bit-level accuracy can waste transmission on redundant information while failing to preserve intended meaning. The introduction distinguishes semantic coding from semantic communication and frames channel variability and semantic objectives as open problems addressed by the paper.

I. INTRODUCTION

  • Bit-error-rate optimization can spend bits on redundant information and still produce a decoded message that misses the intended meaning.The article argues these concerns motivate a semantics-aware view of communication.Source: §I · p. 1
  • Semantic coding seeks efficient or more accurate bit transmission, whereas semantic communication targets meaning or goal execution and needs semantic-level metrics.The introduction says the two concepts are often conflated but are not the same.Source: §I · p. 1
  • Existing systems simplify varying wireless channels to fixed-level AWGN and leave semantic transmission goals or large-scale generalization insufficiently addressed.The paper names these the communication problem and the semantic problem, respectively.Source: §I · pp. 1–2
  • The article proposes adaptive JSNC representation depth for varying channels and sentences, plus semantic-similarity learning that can use task-specific, non-differentiable metrics.The proposed semantic communication system replaces semantics-blind CE or MSE objectives.Source: §I · p. 2

II. SEMANTIC CODING: A CONFIDENCE-BASED JOINT SEMANTICS-NOISE CODING (JSNC) MANNER

JSNC addresses changing channel conditions and makes semantic extraction interpretable through iterative embedding refinement guided by confidence.

A. Overview of the Proposed JSNC Method

  • JSNC refines encoder and decoder embeddings to adapt semantic extraction to sentence and channel variation, while confidence evaluates representation quality and guides refinement.The paper presents this as addressing varying channel situations and making the mechanism interpretable.Source: A. Overview of the Proposed JSNC Method · p. 2

B. Semantic Confidence Mechanism

Semantic confidence determines whether an embedding is ready for downstream processing or needs refinement. Distillation repeats that refinement adaptively, up to a system-specific maximum.

B. Semantic Confidence Mechanism

  • When semantic confidence reaches a predefined threshold, the embedding proceeds downstream; otherwise, the confidence module sends it back for refinement.The mechanism evaluates embeddings in both encoder and decoder.Source: B. Semantic Confidence Mechanism · p. 2

C. Semantic Distillation Mechanism

  • The encoder and decoder distill embeddings repeatedly in a coarse-to-fine process controlled by learnable confidence modules.The figure distinguishes message space from feature space and shows JSNC in both encoder and decoder.Source: Figure 2 · p. 2, C. Semantic Distillation Mechanism · p. 2
  • Distillation stops at the maximum number of steps or when semantic extraction is judged ready, adapting to sentence difficulty and channel state.The maximum step count represents the system's cognition ability.Source: C. Semantic Distillation Mechanism · p. 3

D. Implementation Cost and Computational Complexity

The section weighs JSNC's local computation against repeated distillation and introduces semantic-similarity objectives for communication, including differentiable and non-differentiable settings.

D. Implementation Cost and Computational Complexity

  • Repeated semantic distillation adds computation, but the confidence mechanism costs O(LD), compared with O(LD^2) for LSTM and O(L^2D) for Transformer encoders.The N distillation steps add a linear factor N; unlike N channel transmissions in HARQ, local distillation has O(1) time complexity, and the approach is backward compatible.Source: D. Implementation Cost and… · p. 3
  • Multiple distillation steps bring computational burden, which neural-network pruning, quantization, and knowledge distillation could help simplify.The passage also notes the potential benefit of advanced AI chips.Source: §III · p. 3

III. SEMANTIC COMMUNICATION: SIMILARITY-TARGETED SEMANTICS-PERSEVERING COMMUNICATION (SSC)

  • Because CE and MSE weight every bit, word, or pixel equally, the proposed communication objective minimizes semantic distance or maximizes semantic similarity.The passage applies the argument to languages and other kinds of messages.Source: §III · p. 3
  • The SSC model covers both differentiable and non-differentiable semantic-similarity settings.The passage refers to the model overview in Fig. 3.Source: §III · p. 3

A. Semantic Communication with Differentiable Semantic Similarity (SSC-D)

The section distinguishes semantic communication with differentiable similarity metrics from the harder non-differentiable case. It presents reinforcement learning as a general way to optimize semantic similarity without requiring a differentiable metric.

A. Semantic Communication with Differentiable Semantic Similarity (SSC-D)

  • Differentiable task-specific similarity metrics can back-propagate through the transceiver; masked MSE or CE can target salient regions, and many existing semantic-communication works use this setting.The section contrasts this with its focus on the less-investigated non-differentiable case.Source: A. Semantic Communication with… · p. 3

B. Semantic Communication with Non-differentiable Semantic Similarity (SSC-ND)

  • BLEU is non-differentiable, making standard gradient descent and existing deep-learning JSCC systems intractable for that objective.The authors say non-differentiable cases may be more frequent, making the problem urgent.Source: B. Semantic Communication with… · p. 4
  • Reinforcement learning avoids surrogate-loss redesign across changing environments or similarity metrics by using a scalar semantic-similarity reward, yielding a model- and target-agnostic approach for SSC-D and SSC-ND.The authors describe it as a universal solution, while noting that surrogate losses may still be promising in some cases.Source: B. Semantic Communication with… · p. 4
  • The proposed learning process defines a semantic-similarity score, casts encoding-decoding as reinforcement learning, and optimizes accumulated reward.The decoding process determines which semantic token is generated.Source: B. Semantic Communication with… · p. 4
  • For sentence transmission, decoder state and previously generated words form the recurrent state; for images, the state can be an intermediate decoded image updated through small pixel-value changes.This reframes image regression decoding for an RL setting.Source: B. Semantic Communication with… · p. 4
  • RL is one possible approach and may require minor architecture changes; actor-critic methods are presented as effective and simple to implement for sentence and image tasks.The RL procedure also applies to differentiable rewards, though implementation may require more decoding time.Source: B. Semantic Communication with… · p. 4

C. The Information Overhead

The section separates communication overhead into semantic, recurrence, and system components. It argues that transmitting only semantically necessary information can lower overhead, while reporting better semantic conveyance under matched information and computation overhead.

C. The Information Overhead

  • The overhead comprises information needed for semantic meaning, extra information for perfect bit-level recovery, and content-agnostic system burden.The passage names these the semantics, recurrence, and system parts, respectively.Source: C. The Information Overhead · p. 4
  • Only the semantics part is necessarily needed for transmission, theoretically giving semantic communication lower information overhead; reliable communication is a special case where all bits are treated as necessary.The passage characterizes semantic communication as beyond Shannon’s reliable communication.Source: C. The Information Overhead · p. 4
  • The proposed SSC scheme better conveys semantic meanings than the conventional model under the same information and computation overhead.Semantic extraction itself consumes computational resources and adds local information computations.Source: C. The Information Overhead · p. 4

A. Datasets

The experiments cover sentence, image, and interactive communication, testing confidence-based semantic coding and reinforcement-learning semantic communication across channel conditions. The proposed coding reduces word error, while the communication system learns to preserve meaning and converges stably.

A. Datasets

  • Experiments use European-parliament for sentence transmission, CIFAR-10 for image transmission, and a new “speak something” scenario across AWGN and FIF channels.The authors note that no channel-adaptation technique is used in these simulations.Source: A. Datasets · p. 4

B. Semantic Coding

  • Distillation increases with sentence length and as SNR degrades, generally requiring more rounds on FIF than AWGN, but saturates under extreme noise.The authors interpret these patterns as adaptive exploration of deeper semantic representations; Fig. 4(c-d) reports average distillation times.Source: B. Semantic Coding · p. 5, Figure 4 · p. 5
  • At 10dB AWGN, TF+Conf achieves a 0.038% word error rate, 24% lower than TF and below the conventional 5-bit fixed-length RS code.The reported coding results also show robustness on phase-invariant fading compared with the model without distillation.Source: Figure 4 · p. 5, B. Semantic Coding · p. 5

C. Semantic Communication

  • SSC-ND escapes initial local minima and converges stably on sentence and image tasks, while training remains more challenging on FIF than AWGN.Fig. 5 plots training loss and runtime reward; the loss is shown only for AWGN.Source: C. Semantic Communication · p. 5, Figure 5 · p. 6
  • SSC favors semantically similar words and phrases over bit-level accuracy, preserving key meanings in examples including severely noisy sentences.Table I presents real-life examples of the proposed similarity-targeted communication mechanism.Source: Table I · p. 6, C. Semantic Communication · p. 6

D. Real Communication Scenario: Speak Something

In the “speak something” scenario, SSC scores around 2.3 against the conventional baseline’s 1.2 and requires fewer turns to identify the object. The JSNC improvement is limited and saturates as accuracy approaches 1.

D. Real Communication Scenario: Speak Something

V. OPEN CHALLENGES WITH THE PROPOSED APPROACH

The paper identifies unresolved challenges in semantic information theory, semantic security, and implementation. These include universal theory, user preferences, computation-performance trade-offs, AI infrastructure demands, and task-specific generalization.

V. OPEN CHALLENGES WITH THE PROPOSED APPROACH

  • A universal semantic information theory remains undeveloped, despite progress on practical communication with simple semantic concepts.Source: §V · p. 7
  • Semantic security research remains limited, with user preference and implementation cost among the challenges for semantics-aware systems.Semantic tokens can carry different meanings for different receiver groups and tasks.Source: §V · p. 7
  • Implementation faces computation-performance trade-offs, demands on AI infrastructure, and task-specific generalization challenges.Source: §V · p. 7

VI. CONCLUSION AND OUTLOOK

The article reframes semantics-aware communication beyond semantic coding, combining an overview and system-design outlook with JSNC and RL-based semantic communication mechanisms. It presents these mechanisms as demonstrating feasibility and promising benefits, while noting that semantic-communication work remains sparse and restricted to differentiable settings.

VI. CONCLUSION AND OUTLOOK

  • The article surveys existing semantics-aware techniques, analyzes their drawbacks, and rethinks system design, including implementation cost and future designs.Source: §VI · p. 7
  • JSNC is proposed for semantic coding, while SSC-D and SSC-ND use RL-based similarity-targeted semantic communication for differentiable and non-differentiable metrics.Together, these mechanisms demonstrate feasibility and promising benefits.Source: §VI · p. 7
  • Most existing semantics-aware work falls under semantic coding; semantic communication has few studies and is restricted to differentiable settings.Source: §VI · p. 7
Loading 2110.08496v3…